espnet2.enh.layers.dcunet.FeatureMapDense

source

class espnet2.enh.layers.dcunet.FeatureMapDense(input_dim, output_dim, complex_valued=False)

Bases: Module

A fully connected layer that reshapes outputs to feature maps.

This layer is designed to be used in the context of complex-valued neural networks, where it applies a dense linear transformation to the input and reshapes the output to add two additional dimensions for feature maps. It utilizes the ComplexLinear layer to handle complex-valued inputs.

complex_valued

Indicates if the layer processes complex-valued inputs.

• Type: bool

dense

The underlying dense layer that performs the linear transformation.

• Type:ComplexLinear

• Parameters:

  • input_dim (int) – The number of input features.
  • output_dim (int) – The number of output features.
  • complex_valued (bool) – Whether the layer should support complex values. Defaults to False.

• Returns: The reshaped output tensor with added dimensions for feature maps.

• Return type: Tensor

####### Examples

>>> layer = FeatureMapDense(input_dim=128, output_dim=64, complex_valued=True)
>>> input_tensor = torch.randn(10, 128) + 1j * torch.randn(10, 128)
>>> output_tensor = layer(input_tensor)
>>> output_tensor.shape
torch.Size([10, 64, 1, 1])

NOTE

This layer is primarily intended for use in complex-valued models, such as those used for audio signal processing or speech enhancement.

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x)

A fully connected layer that reshapes outputs to feature maps.

This layer takes an input tensor and transforms it into a shape suitable for subsequent operations in a neural network. It is designed to handle both real and complex-valued inputs, depending on the configuration.

complex_valued

Indicates whether the layer operates on complex numbers.

• Type: bool

dense

A linear layer that handles the transformation of input dimensions.

• Type:ComplexLinear

• Parameters:

  • input_dim (int) – The number of input features.
  • output_dim (int) – The number of output features.
  • complex_valued (bool , optional) – If True, the layer will handle complex inputs. Defaults to False.

• Returns: The reshaped output tensor.

• Return type: Tensor

####### Examples

>>> import torch
>>> layer = FeatureMapDense(input_dim=128, output_dim=256)
>>> input_tensor = torch.randn(10, 128)  # Batch of 10 samples
>>> output_tensor = layer(input_tensor)
>>> print(output_tensor.shape)  # Should print (10, 256, 1, 1)

NOTE

The output tensor will have two additional dimensions (1, 1) added to facilitate reshaping into feature maps for subsequent layers.